Claude-skill-registry image

ImageAgent - Generate EVOLEA Brand Images with Reinforcement Learning

install
source · Clone the upstream repo
git clone https://github.com/majiayu000/claude-skill-registry
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/majiayu000/claude-skill-registry "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/data/image-generation-rl" ~/.claude/skills/majiayu000-claude-skill-registry-image && rm -rf "$T"
manifest: skills/data/image-generation-rl/SKILL.md
source content

EVOLEA Image Generation with Reinforcement Learning

Invoke with:

/image
or use automatically when generating images for the website Last Updated: 2025-12-27

This skill provides an intelligent image generation system that learns from user feedback to produce increasingly better, brand-consistent images for EVOLEA.


MCP Server (Claude Desktop / Claude App)

The image generation system is available as an MCP server for use with Claude Desktop and Claude App.

Available Tools

ToolDescription
generate_image
Generate a single image with EVOLEA branding
generate_ab_comparison
Generate A/B comparison for training (2 options + grid)
list_generated_images
List recent generated images by category
get_brand_prompt_template
Get prompt templates for specific image types
publish_image
Publish image to GitHub for public URL access
get_training_guide
Get the full training guide documentation

Setup (Claude Desktop)

Add to

%APPDATA%\Claude\claude_desktop_config.json
:

{
  "mcpServers": {
    "evolea-images": {
      "command": "python",
      "args": ["C:/Users/christoph/evolea-website/scripts/mcp_image_server.py"],
      "env": {
        "GOOGLE_API_KEY": "your-gemini-api-key"
      }
    }
  }
}

Quick Commands in Claude Desktop

"Generate an image of children doing art together"
"Create an A/B comparison for Mini Projekte hero"
"Show me recent training images"
"Publish the image to GitHub so I can see it on my phone"

How It Works

┌─────────────────────────────────────────────────────────────┐
│  1. LOAD LEARNINGS from LEARNINGS.md                        │
│     - Apply positive patterns (+3, +1 boosts)               │
│     - Include negative patterns in exclusions (-1, -3)      │
├─────────────────────────────────────────────────────────────┤
│  2. GENERATE OPTIONS (A and B)                              │
│     - Option A: Base prompt with current learnings          │
│     - Option B: Base + additional style modifiers           │
│     - Save to: public/images/generated/training/            │
├─────────────────────────────────────────────────────────────┤
│  3. CREATE COMPARISON GRID                                  │
│     - Side-by-side A|B image for easy comparison            │
├─────────────────────────────────────────────────────────────┤
│  4. USER SELECTS                                            │
│     - A, B, or Neither with feedback                        │
├─────────────────────────────────────────────────────────────┤
│  5. UPDATE LEARNINGS                                        │
│     - Boost winning patterns                                │
│     - Record negative patterns from feedback                │
│     - Persist to LEARNINGS.md                               │
├─────────────────────────────────────────────────────────────┤
│  6. PUBLISH (optional)                                      │
│     - Push to GitHub → Cloudflare Pages                     │
│     - Get public URL for mobile access                      │
├─────────────────────────────────────────────────────────────┤
│  7. ITERATE until satisfied                                 │
└─────────────────────────────────────────────────────────────┘

Command Line Usage

Generate with Training Loop

# Start a training session for a specific image
python scripts/generate_image.py "children ages 5-8 creating art" \
  --name mini-projekte-hero \
  --category programs \
  --training

Manual A/B Generation

# Generate 2 options with comparison grid
python scripts/generate_image.py "your prompt" \
  --name image-name \
  --count 2 \
  --comparison-grid

Fully Automated (with Claude selection)

python scripts/generate_image.py "your prompt" \
  --name image-name \
  --auto-select

Key Files

FilePurpose
scripts/mcp_image_server.py
MCP server for Claude Desktop
scripts/generate_image.py
Core image generation script
LEARNINGS.md
Persistent style preferences with scores
TRAINING-GUIDE.md
User guide for iterative training
training-log.json
Full history of all training sessions
style-profiles/*.md
Pre-defined style templates

EVOLEA Style Guidelines

Central European Children (Ages 5-8)

  • Skin tones: Light/fair with warm undertones
  • Hair colors: Blonde, light brown, auburn, brown
  • Features: Soft, rounded, friendly expressions
  • NO: Religious symbols, American cultural elements

Whimsical Style (User's Preferred)

  • Soft dreamy watercolor textures
  • Floating, layered clouds in spectrum colors
  • Ethereal, magical atmosphere
  • Delicate butterflies or unicorns
  • Pastel gradient backgrounds (lavender, mint, coral, cream)

Color Palette

ColorHexUsage
Magenta#DD48E0Primary accents
Mint#7BEDD5Fresh, nature
Lavender#CD87F8Calm, creative
Coral#FF7E5DEnergy, warmth
Cream#FFFBF7Backgrounds

NEVER Include

  • Puzzle piece symbols
  • Clinical/medical settings
  • Religious symbols on children
  • American cultural elements (yellow school buses)
  • Photorealistic style
  • Dark or muted colors

Teacher Illustrations

Based on EVOLEA team members:

NameDescription
Gianna SpiessFemale, professional, warm, M.Sc. BCBA
Annemarie EliasFemale, friendly, approachable, M.Sc. BCBA
Christoph JennyMale, supportive, engaged
Alexandra AleksicFemale, young, energetic, B.Sc.

Prompt Templates

Program Hero

Children aged [AGE] [ACTIVITY] in [SETTING].
Swiss/Central European children with light skin and varied natural hair colors.
[SPECIFIC DETAILS].
Mood: [EMOTION]. Dominant colors: [COLORS].
Soft watercolor children's book illustration style with delicate butterflies.

Abstract/Decorative

Soft dreamy [SUBJECT] in whimsical watercolor style.
Layered colorful clouds in lavender, mint, coral, and cream.
Ethereal atmosphere with gentle sparkles.
Delicate butterflies floating softly.
Dominant colors: [COLORS]. Mood: magical, warm, inviting.

Backend Configuration

The image generation always uses Gemini 3 Pro:

ModelID
Gemini 3 Pro
gemini-3-pro-image-preview

This is configured in

scripts/generate_image.py
via
CONFIG.gemini_model
.

Replicate (Flux) is available as a fallback if Gemini is unavailable in your region.


Related Files

  • TRAINING-GUIDE.md
    - Detailed training workflow guide
  • LEARNINGS.md
    - Current learned preferences
  • style-profiles/
    - Style profile templates
  • .claude/skills/Design skills/illustrations.md
    - Illustration guidelines